本研究旨在提出架構以心理特徵（數值）結合文本應用深度學習預測網路霸凌行為，不同於過去研究只由文本或心理特徵（數值）的單一方式訓練模型進行預測，本研究擬調用IBM API分析網路貼文中的情緒、情感以及發文者的五型人格傾向。文本方面，本研究使用的詞嵌入模型有GloVe與Word2Vec，並依據實驗結果提出較適合本研究的詞嵌入模型。 在文本結合心理特徵（數值）輸入雙向長短期記憶神經網路串接全連接神經網路的部分則較單純使用心理特徵（數值）輸入全連接神經網路在各項指標中更佳，在epoch （訓練期數）本研究實驗10、20與30的組合，在epoch為20的情況下在各項指標的結果最佳。比較epoch同為20的狀況下loss值由0.3673下降至0.2842，F1 score提升近6% 。除了進行網路霸凌行為預測外，本研究亦探討各個心理特徵與網路霸凌行為間的相關性。 本研究結果可提供社群媒體網站一個預測機制，應用於預警或篩選出含有網路霸凌成分的貼文，並提供給社群媒體平台經營方，即早覺察在文本中顯露出的相關負面人格、情緒與情感。避免惡意傷害的貼文對網路使用者造成傷害。 ;This research aims to propose a new method to predict cyberbullying behavior by combining psychology features with text. Each of the many posts on the Internet contains their own emotions, personalities and their sentiments at the moment of posting. These emotions, sentiments and personalities can be identified through applications. In terms of text, the Word Embedding methods used in this study are GloVe and Word2Vec. We compared them and proposed the methods which more suitable for this study. It is different from the past research that only used a one-way information method of text or psychology features values to train models to make predictions. The use of text and psychological features, in contrast of purely using numerical data, loss value drops from 0.3673 to 0.2842 and F1 score increase by nearly 5% under the same epoch. This study in the epoch (Number of training periods) has experimented with a combination of 10, 20, and 30, the results of each metrics are the best when the epoch is 20. In addition to predicting cyberbullying, this research also explores the correlation between various psychological features and cyberbullying. The results of this research confirmed our proposed approach to be a useful detection mechanism for social media websites to provide an early warning or to screen out maliciously posts occurred in Internet.
